NJIT's EarlyDetect Forecasts Solar Active Regions Nine Hours Before Sunspots Appear, Using a Transformer on Helioseismic Data
A study published 2026-08-14 in the Journal of Geophysical Research: Machine Learning and Computation describes EarlyDetect, a Transformer model that predicts the emergence of solar active regions an average of nearly nine hours in advance by reading acoustic power maps and magnetic field measurements from NASA's Solar Dynamics Observatory HMI instrument, sampled every 45 seconds. The corresponding author is NJIT undergraduate Jonas Tirona, working with NJIT computer scientists and solar physicists plus collaborators at Princeton and NASA Ames. Notably, the team found that a data-filtering step they had assumed was helping was actually degrading forecasts — a reminder that preprocessing intuitions transfer badly to Transformer architectures.
Source
↳ Follow the thread